Why Now Progress How It Works ASR Demo Speakers What We Need
Community-Driven Language Preservation

Kinspeak

Keeping Anindilyakwa alive — building speech and language tools with, and for, the community of Groote Eylandt. Every voice recorded, every word preserved, stays under community ownership.

Listen to the Language ↓
📍 Groote Eylandt, Northern Territory, Australia
Why Now

A Closing Window

The Window
The eldest fluent Anindilyakwa speakers are in their 70s and 80s. Every year without recording is knowledge permanently lost — pronunciations, stories, and ways of speaking that no dictionary can capture.
🔬
The Shift
For the first time, AI can learn an endangered language from just hours of recordings. Five years ago this required thousands of hours. The technology is ready — but it needs voices to learn from.
🌏
The Opportunity
Anindilyakwa can be the model for every at-risk language in Australia. What we build here — the tools, the process, the data sovereignty framework — becomes the blueprint others follow.
"We have a narrow window. The technology is ready.
The question is whether we act in time."
About

About Kinspeak

This project was started by Rodney Woods, who grew up on Groote Eylandt and has family ties to the Anindilyakwa community. Kinspeak is not a company — it's an open effort to build the tools that keep this language alive. Every piece of technology we create is designed to be owned and controlled by the community.

Progress

Where We Are

What we've built so far — and what's needed to reach full coverage.

Recorded Speech
80h of 500h+ needed
Natural speech + validated synthetic data from PARADISEC archives
Voices Preserved
4 of 30+ fluent elders
Each new speaker adds unique vocabulary and pronunciation
Words Documented
4,200 of ~15,000 estimated
Hand-built lexicon with morphological analysis
Transcription Accuracy
95% — target 98%+
Already approaching human-level accuracy on known speakers
How It Works

What This Technology Does

In plain terms — here is what each part of the system does and why it matters for language preservation.

🎙 → 📝
Listening & Transcribing
We teach a computer to listen to Anindilyakwa speech and write down what it hears — like a very patient transcriber that never gets tired. This is what linguists call speech recognition.
📝 → 🔊
Reading & Speaking
We teach a computer to read Anindilyakwa text aloud, in the voice of real speakers — so elders' voices can be heard by future generations. This is called voice synthesis.
🔤 ↔ 🔤
Translating
We build a bridge between Anindilyakwa and English — word by word, learning the language's rich structure where a single word can carry the meaning of an entire English sentence.
📖 What do the numbers mean?

Think of it like proofreading — the lower the error number, the better the computer listened.

CER — Character Error Rate
How many individual letters the computer got wrong when transcribing speech. 0% = perfect, it heard every letter correctly.
WER — Word Error Rate
How many whole words were wrong. A stricter measure — getting one letter wrong in a word counts the entire word as an error.
Transcription Accuracy
The flip side of error rate. When we say "100% accuracy" on our best samples, it means the computer perfectly matched what the human speaker said.
Live Demo
Speech Recognition in Action
Press the button to hear a real Anindilyakwa recording — watch the computer transcribe it in real-time
What the computer heard
···
What was actually said
Hear the Language

Speaker Showcase

Listen to natural and AI-synthesized Anindilyakwa speech. Each sample shows the reference text and ASR transcription with character error rate (CER).

What We Need

How ALC Can Help

The technology works. To take it further, we need the community's support.

🎙
Access to Speakers
We need to sit with more elders and record their speech — stories, conversations, everyday language. Each speaker adds a unique voice and vocabulary that strengthens every tool we build.
🤝
Partnership & Endorsement
A formal partnership with ALC ensures the community controls this work. Data sovereignty means every recording, every model, every tool belongs to the community — not to us, not to a university.
💬
Community Input
Which words matter most? Which stories should be preserved first? The community decides what we prioritise. The technology serves the language — not the other way around.
To discuss a partnership, reach out to Rodney Woods
Try It Live

Translation Demo

Bidirectional machine translation between English and Anindilyakwa, powered by morphological analysis and a growing lexicon.

English
Anindilyakwa
⚡ Server Offline

The translation server isn't running right now.
Start it with python translation/translation_server.py

Under the Hood

Our Technology Stack

A custom pipeline built for low-resource language preservation, combining state-of-the-art models with domain-specific engineering.

🎙
Speech Recognition
A computer that has learned to listen to Anindilyakwa and write down what it hears. Trained on recordings from the PARADISEC cultural archive, it can now transcribe speech with near-perfect accuracy.
🗣
Voice Preservation
Technology that can read Anindilyakwa text aloud in the voice of real speakers. Elders' voices can be preserved and heard by future generations, even when they are no longer with us.
🔤
Translation Engine
A growing bridge between Anindilyakwa and English, built word by word. It understands the language's rich system of prefixes and suffixes — where a single word can carry the meaning of an entire English sentence.
0:00